联合各个“部落”:用文本获取营销洞察 · Conceptual review and methods guide (no new data) 概念综述与方法指南(无新数据) · Ch 2, 8, 13 · 35 min
Berger, J., Humphreys, A., Ludwig, S., Moe, W. W., Netzer, O., & Schweidel, D. A. (2020). Uniting the tribes: Using text for marketing insight. Journal of Marketing, 84(1), 1–25. https://doi.org/10.1177/0022242919873106
| Pages 页码 | What to read | 读什么 |
|---|---|---|
| 1–2 | Abstract and the four contributions (3 min). | 摘要与四个贡献(3 分钟)。 |
| 7–8 | The 4-step workflow and Table 2 (5 min). | 4 步流程与表 2(5 分钟)。 |
| 10–12 | Table 4 and the dictionary, topic model and relation sections (6 min). | 表 4,以及词典、主题模型和关系抽取部分(6 分钟)。 |
| 13–16 | Validity section and Table 5, our method checklist (6 min). | 效度部分与表 5,即我们的方法检查清单(6 分钟)。 |
Pages · 页码 1–2
Pages · 页码 2–4
Pages · 页码 4–5
Pages · 页码 5
Pages · 页码 5–6
Pages · 页码 6–7
Pages · 页码 7–9
Pages · 页码 9–12
Pages · 页码 12–13
Pages · 页码 13–15
Pages · 页码 15–16
Pages · 页码 16–18
Pages · 页码 18–19
Pages · 页码 19–20
Pages · 页码 21
| Exhibit 图表 | How to read it | 怎么看 |
|---|---|---|
| Table 1 (p. 3) | A grid of text producers (rows) by receivers (columns): consumers, firms, investors, society. Find our data in the consumer-to-consumer cell: reviews and social media. | 以文本生产者(行)和接收者(列)构成的矩阵:消费者、企业、投资者、社会。我们的数据位于“消费者对消费者”格:评论和社交媒体。 |
| Table 2 (p. 8) | The full workflow in four columns: preprocessing, common tools, measurement, validity. Read it left to right as a checklist. | 四栏完整流程:预处理、常用工具、测量、效度。从左到右当作检查清单使用。 |
| Table 3 (p. 9) | Each preprocessing step with the issues to consider and a Twitter example. Useful for writing our cleaning steps. | 列出每个预处理步骤的注意事项,并配有 Twitter 例子。可用于撰写我们的数据清洗步骤。 |
| Table 4 (p. 10) | Taxonomy of entity, topic and relation extraction: tools, questions, benefits, limits and examples. The limits column is the key part. | 实体、主题和关系抽取的分类:工具、研究问题、优点、局限和实例。“局限”一栏最关键。 |
| Table 5 (p. 14) | Validation techniques by validity type, with the method and a reference. It gives the 10% sample, 80% hit rate and 80%–90% holdout rules. | 按效度类型列出检验方法和参考文献。其中给出 10% 抽样、80% 命中率和 80%–90% 留出样本等规则。 |
| Term 术语 | Meaning | 解释 |
|---|---|---|
| automated text analysis 自动文本分析 | Using computers to turn text into measurable data. | 用计算机把文本转化为可测量的数据。 |
| text reflects vs impacts 文本的反映与影响 | Text says something about its writer, and it also changes its reader. | 文本既说明写作者的情况,也会改变阅读者。 |
| dictionary (lexicon) 词典(词表) | A fixed word list that assigns words to categories such as positive or negative. | 一份固定的词表,把词语归入正面、负面等类别。 |
| bag of words 词袋 | Treating text as a pile of words and ignoring their order. | 把文本当作一堆词,忽略词语的顺序。 |
| entity extraction 实体抽取 | Finding and counting single words or phrases such as brand names. | 找出并统计品牌名等单个词或短语。 |
| topic model (LDA) 主题模型(潜在狄利克雷分配) | A statistical method that finds groups of words that tend to appear together. | 一种统计方法,找出经常一起出现的词群。 |
| relation extraction 关系抽取 | Identifying how entities link, such as which feature has which problem. | 识别实体之间的联系,例如哪个功能有哪个问题。 |
| recall and precision 召回率与精确率 | Recall: how many true cases were found. Precision: how many found cases were right. | 召回率:找到了多少真实个案。精确率:找到的个案有多少是对的。 |
| construct validity 构念效度 | Whether the text measure really captures the concept we claim. | 文本指标是否真正测到了我们声称的概念。 |
| hit rate / saturation 命中率/饱和法 | Checking a sample of coded cases by hand and refining the word list until enough are correct. | 人工检查一部分已编码样本,并修改词表,直到正确比例足够高。 |
| holdout sample / k-fold 留出样本/k 折交叉验证 | Training a model on part of the data and testing it on the rest. | 用部分数据训练模型,再用剩余数据测试。 |
| self-selection bias 自我选择偏差 | Bias because only some people choose to write reviews or posts. | 因为只有部分人选择写评论或发帖而产生的偏差。 |
| Value 数值 | What it measures | 含义 | Page 页 |
|---|---|---|---|
| 80%–95% | Share of business data estimated to be unstructured, mostly text | 估计为非结构化数据(主要是文本)的商业数据比例:80%–95% | p. 1 |
| 84(1), 1–25 | Journal of Marketing volume, issue and pages | 《Journal of Marketing》卷期与页码:84(1), 1–25 | p. 1 |
| 280 characters | Tweet length limit, an example of a platform constraint on text | 推文长度上限 280 个字符,是平台约束文本的例子 | p. 5 |
| 4 steps | Number of steps in the text analysis process; step 4 is validity | 文本分析流程共 4 步(4 steps);第 4 步是效度检验 | p. 7 |
| Table 4 | Taxonomy of text analysis tools | 表 4:文本分析工具分类 | p. 10 |
| EL 2.0 | One of the common emotion dictionaries named | 文中列出的常用情绪词典之一:EL 2.0 | p. 11 |
| 2003 | Year of the original LDA topic model (Blei, Ng and Jordan) | 原始 LDA 主题模型的发表年份(Blei、Ng 和 Jordan):2003 | p. 11 |
| 1–12 | Readability scale, matching U.S. school grade levels | 可读性评分范围 1–12,对应美国学校年级 | p. 13 |
| 80% | Target hit rate when validating a dictionary by saturation (Weber 2005) | 用饱和法检验词典时的目标命中率 80%(Weber 2005) | p. 13 |
| 10% | Share of coded cases pulled to check a dictionary | 检验词典时抽取的已编码样本比例:10% | p. 14 |
| 80%–90% | Share of data used to train a model, testing on the rest | 用于训练模型的数据比例 80%–90%,其余用于测试 | p. 14 |
| Table 5 | Text analysis validation techniques | 表 5:文本分析效度检验方法 | p. 14 |
| 2012 | Tirunillai and Tellis: Twitter volume affects stock price; only negative sentiment predicts | Tirunillai 和 Tellis(2012):推特数量影响股价;只有负面情感有预测力 | p. 15 |
| 2018 | Smith and Anderson: Twitter users tend to be younger and more educated | Smith 和 Anderson(2018):推特用户偏年轻、学历偏高 | p. 16 |
Written by the group from the full text. Numbers and page references were checked against the source. Licensed cases and reports are summarised, not reproduced.
由小组根据全文整理,数字与页码已对照原文核对。授权案例与报告只作摘要,不转载原文。